Tool Execution

Overview

When an LLM returns tool calls in its response, Bifrost does not automatically execute them. Instead, your application explicitly calls the tool execution API, giving you full control over:

Which tool calls to execute
User approval workflows
Security validation
Audit logging

The basic flow is: Chat Request → Review Tool Calls → Execute Tools → Continue Conversation. For detailed architecture diagrams, see the MCP Architecture documentation.

Authentication

The /v1/mcp/tool/execute endpoint uses the same authentication as other inference endpoints like /v1/chat/completions:

Auth Configuration	Behavior
`disable_auth_on_inference: true`	No auth required
`disable_auth_on_inference: false`	Auth required

Virtual keys and authentication are independent layers that work together. For details on how to use virtual keys with authentication, see Authentication and Virtual Keys.

End-to-End Example

Gateway
Go SDK

Step 1: Send Chat Request

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o",
    "messages": [
      {
        "role": "user",
        "content": "List files in the current directory"
      }
    ]
  }'

Response with tool calls:

{
  "id": "chatcmpl-abc123",
  "choices": [{
    "index": 0,
    "message": {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "call_xyz789",
        "type": "function",
        "function": {
          "name": "filesystem_list_directory",
          "arguments": "{\"path\": \".\"}"
        }
      }]
    },
    "finish_reason": "tool_calls"
  }]
}

Tool names are prefixed with the MCP client name (e.g., filesystem_list_directory). This ensures uniqueness across multiple MCP clients.

Step 2: Execute the Tool

The request body matches the tool call object from the response:

curl -X POST http://localhost:8080/v1/mcp/tool/execute \
  -H "Content-Type: application/json" \
  -d '{
    "id": "call_xyz789",
    "type": "function",
    "function": {
      "name": "filesystem_list_directory",
      "arguments": "{\"path\": \".\"}"
    }
  }'

Tool result response:

{
  "role": "tool",
  "content": "[\"config.json\", \"main.go\", \"README.md\"]",
  "tool_call_id": "call_xyz789"
}

Step 3: Continue the Conversation

Assemble the full conversation history and continue:

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o",
    "messages": [
      {
        "role": "user",
        "content": "List files in the current directory"
      },
      {
        "role": "assistant",
        "content": null,
        "tool_calls": [{
          "id": "call_xyz789",
          "type": "function",
          "function": {
            "name": "filesystem_list_directory",
            "arguments": "{\"path\": \".\"}"
          }
        }]
      },
      {
        "role": "tool",
        "content": "[\"config.json\", \"main.go\", \"README.md\"]",
        "tool_call_id": "call_xyz789"
      }
    ]
  }'

Final response:

{
  "choices": [{
    "message": {
      "role": "assistant",
      "content": "The current directory contains 3 files:\n\n1. **config.json** - Configuration file\n2. **main.go** - Go source file\n3. **README.md** - Documentation"
    },
    "finish_reason": "stop"
  }]
}

package main

import (
    "context"
    "fmt"
    bifrost "github.com/maximhq/bifrost/core"
    "github.com/maximhq/bifrost/core/schemas"
)

func main() {
    // Initialize Bifrost with MCP (see Connecting to Servers)
    client, _ := bifrost.Init(context.Background(), config)

    // Step 1: Send initial request
    firstMessage := schemas.ChatMessage{
        Role: schemas.ChatMessageRoleUser,
        Content: schemas.ChatMessageContent{
            ContentStr: bifrost.Ptr("List files in the current directory"),
        },
    }

    request := &schemas.BifrostChatRequest{
        Provider: schemas.OpenAI,
        Model:    "gpt-4o",
        Input:    []schemas.ChatMessage{firstMessage},
    }

    response, err := client.ChatCompletionRequest(schemas.NewBifrostContext(context.Background(), schemas.NoDeadline), request)
    if err != nil {
        panic(err)
    }

    // Build conversation history
    history := []schemas.ChatMessage{firstMessage}

    // Step 2: Process tool calls
    if response.Choices[0].Message.ToolCalls != nil {
        assistantMessage := response.Choices[0].Message
        history = append(history, assistantMessage)

        for _, toolCall := range *assistantMessage.ToolCalls {
            fmt.Printf("Tool requested: %s\n", *toolCall.Function.Name)

            // YOUR APPROVAL LOGIC HERE
            // - Validate arguments
            // - Check permissions
            // - Get user confirmation if needed

            // Step 3: Execute the tool
            toolResult, err := client.ExecuteChatMCPTool(context.Background(), toolCall)
            if err != nil {
                fmt.Printf("Tool execution failed: %v\n", err)
                continue
            }

            fmt.Printf("Tool result: %s\n", *toolResult.Content.ContentStr)
            history = append(history, *toolResult)
        }
    }

    // Step 4: Continue conversation with results
    finalRequest := &schemas.BifrostChatRequest{
        Provider: schemas.OpenAI,
        Model:    "gpt-4o",
        Input:    history,
    }

    finalResponse, err := client.ChatCompletionRequest(schemas.NewBifrostContext(context.Background(), schemas.NoDeadline), finalRequest)
    if err != nil {
        panic(err)
    }

    fmt.Printf("Final response: %s\n", *finalResponse.Choices[0].Message.Content.ContentStr)
}

Response Formats

Bifrost supports two API formats for tool execution:

Chat Format (Default)

Use ?format=chat or omit the parameter:

POST /v1/mcp/tool/execute?format=chat

Request:

{
  "id": "call_xyz789",
  "type": "function",
  "function": {
    "name": "filesystem_read_file",
    "arguments": "{\"path\": \"config.json\"}"
  }
}

Response:

{
  "role": "tool",
  "content": "{\"key\": \"value\"}",
  "tool_call_id": "call_xyz789"
}

Responses Format

Use ?format=responses for the Responses API format:

POST /v1/mcp/tool/execute?format=responses

Request:

{
  "type": "function_call_output",
  "call_id": "call_xyz789",
  "name": "filesystem_read_file",
  "arguments": "{\"path\": \"config.json\"}"
}

Response:

{
  "type": "function_call_output",
  "call_id": "call_xyz789",
  "output": "{\"key\": \"value\"}"
}

Multiple Tool Calls

LLMs often request multiple tools in a single response. Execute them in sequence or parallel:

Sequential
Parallel

for _, toolCall := range *response.Choices[0].Message.ToolCalls {
    result, err := client.ExecuteChatMCPTool(ctx, toolCall)
    if err != nil {
        // Handle error
        continue
    }
    history = append(history, *result)
}

toolCalls := *response.Choices[0].Message.ToolCalls
results := make([]*schemas.ChatMessage, len(toolCalls))
var wg sync.WaitGroup

for i, toolCall := range toolCalls {
    wg.Add(1)
    go func(idx int, tc schemas.ChatAssistantMessageToolCall) {
        defer wg.Done()
        result, err := client.ExecuteChatMCPTool(ctx, tc)
        if err == nil {
            results[idx] = result
        }
    }(i, toolCall)
}

wg.Wait()

for _, result := range results {
    if result != nil {
        history = append(history, *result)
    }
}

Error Handling

Tool execution can fail for various reasons:

result, err := client.ExecuteChatMCPTool(ctx, toolCall)
if err != nil {
    switch {
    case errors.Is(err, context.DeadlineExceeded):
        // Tool execution timed out
    case strings.Contains(err.Error(), "tool not found"):
        // Tool doesn't exist or client disconnected
    case strings.Contains(err.Error(), "not allowed"):
        // Tool filtered out by configuration
    default:
        // Other execution error
    }
}

Gateway error responses:

{
  "error": {
    "type": "tool_execution_error",
    "message": "Tool 'filesystem_delete_file' is not allowed for this request"
  }
}

Auth failure recovery

Inline auth-failure retry is available in Bifrost v2.0.0 and above.

When a tool call reaches the upstream MCP server but comes back with a clean auth rejection (401/403), Bifrost reacts on the spot instead of waiting for the health monitor:

Per-user clients (per_user_oauth, per_user_headers): Bifrost force-refreshes the caller’s credential, re-acquires a connection, and retries the same call inline, exactly once. If the retry succeeds, the caller just sees a normal success.
Shared clients (oauth, headers): Bifrost immediately force-refreshes the credential and starts reconnecting in the background. The failing call waits briefly for that reconnect (capped by its own deadline) and retries once on the healed connection; the caller usually sees a normal success. If the reconnect doesn’t finish in time, the original error surfaces and later calls succeed once the background reconnect completes.
Safety opt-out: tools annotated as destructive and not idempotent are never retried; the reconnect still happens, but the original error always surfaces so a side effect cannot run twice.

This applies to POST /v1/mcp/tool/execute, Agent Mode tool calls, and tool calls over the MCP Gateway alike.

Copy-Pastable Responses

Tool execution responses are designed to be directly appended to your conversation history:

// Tool result is already in the correct format
toolResult, _ := client.ExecuteChatMCPTool(ctx, toolCall)

// Just append it directly
history = append(history, *toolResult)

The response includes:

Correct role field ("tool")
Matching tool_call_id for correlation
Properly formatted content

Next Steps

Agent Mode

Enable autonomous tool execution with auto-approval

Tool Filtering

Control which tools are available per request

Overview

Quick Start

Release Cadence

Migration Guides

SDK Integrations

Providers & Guides

MCP Gateway

Custom plugins

Open Source Features

Overview

Authentication

End-to-End Example

Step 1: Send Chat Request

Step 2: Execute the Tool

Step 3: Continue the Conversation

Response Formats

Chat Format (Default)

Responses Format

Multiple Tool Calls

Error Handling

Auth failure recovery

Copy-Pastable Responses

Next Steps

Agent Mode

Tool Filtering

​Overview

​Authentication

​End-to-End Example

​Step 1: Send Chat Request

​Step 2: Execute the Tool

​Step 3: Continue the Conversation

​Response Formats

​Chat Format (Default)

​Responses Format

​Multiple Tool Calls

​Error Handling

​Auth failure recovery

​Copy-Pastable Responses

​Next Steps

Agent Mode

Tool Filtering

Overview

Authentication

End-to-End Example

Step 1: Send Chat Request

Step 2: Execute the Tool

Step 3: Continue the Conversation

Response Formats

Chat Format (Default)

Responses Format

Multiple Tool Calls

Error Handling

Auth failure recovery

Copy-Pastable Responses

Next Steps